Kredal
Eight Gemini agents that check financial documents before submission and ask a person when they're unsure.
- Gemini
- Google ADK
- Cloud Run
- Vertex AI
- Firestore
- Financial documentsGenerated sample files
- Passes the checksScored against 32 rules
- Asks a personBelow the confidence bar
Kredal is a multi-agent system for checking financial documents before they go to a bank. Eight Gemini agents on Google ADK work through them and hand anything uncertain to a person.
What it does
It runs on Cloud Run, Vertex AI and Firestore, with 32 published scoring rules. Confidence bars of 60%, 85% and 97%, depending on severity, mark the point below which it asks a person.
What I measured
92 unit tests pass (6 live tests are skipped). I built it for the AI Builder Cup in Oct 2026 as team lead.
On 9 Nov 2026 Claude ran it on 4 generated sample companies. All four passed, one with an injected model failure that the Healer agent recovered from. It marked an edited bank statement "do not submit" and matched 105 of 105 extracted fields to the known answers. Runs took about a minute.
What it doesn't show
The documents are generated specimens, not customers. No bank has approved anything with it. The results screen still contradicts itself in places, so there are no screenshots or live link here until I fix that.
Questions about Kredal
What is Kredal?
A multi-agent system that checks a small business's financial documents before they go to a bank. Eight Gemini agents on Google ADK work through 32 published scoring rules.
When does Kredal ask a person?
Below a confidence bar of 60%, 85% or 97%, depending on how serious the check is, it hands the case to a person instead of deciding.
How was Kredal tested?
92 unit tests pass. On 9 Nov 2026 it ran on 4 generated sample companies and matched 105 of 105 extracted fields to the known answers.
What doesn't Kredal show?
The documents are generated specimens, not customers, and no bank has approved anything. I built it for the AI Builder Cup in Oct 2026.
Last updated 12 Nov 2026